Tinker Review (2026): Pricing, Features & Honest Verdict

Tinker is Thinking Machines Lab’s training API for fine-tuning open-source models with LoRA. Four functions, published per-token pricing, and none of the cluster management. Best for: ML researchers and developers who want algorithmic control without infrastructure. Price: usage-based per million tokens, from $0.18 (free plan: no). Rating: 8/10.

What is Tinker?

Tinker gives you four functions and takes the GPU cluster off your hands. forward_backward runs a forward and backward pass and accumulates the gradient. optim_step updates the weights. sample generates tokens for interaction, evaluation or reinforcement learning actions. save_state checkpoints so you can resume. That is the whole API surface, and the deliberate smallness of it is the design.

Everything underneath, scheduling, distributed training across GPU clusters, tuning, resource management and reliability, is Thinking Machines’ problem. You bring a dataset of supervised examples or an RL environment, pick a base model, and write the training loop yourself. Tinker uses LoRA, which fine-tunes a small adapter rather than updating all base weights, and the lab’s own research argues that with the right setup LoRA matches full fine-tuning while needing less compute.

We came to it sideways. Inkling, Thinking Machines’ 975B open-weights multimodal model, launched at number 6 on the 20 July Product Hunt board. Tinker is the interaction layer that sits alongside it, it has its own product page and its own published price list, and it is a distinctly different product from the model itself.

What Are Tinker’s Key Features?

Four primitives instead of a training framework

Most managed fine-tuning products give you a form: upload a JSONL, pick a learning rate, press go. Tinker gives you gradient-level control and expects you to write the algorithm. That is a narrower audience and a much higher ceiling. If your research involves a non-standard training setup, this is close to the only hosted option that does not fight you.

A broad model roster you do not have to host

You can train Inkling and Inkling-Small, DeepSeek-V3.1, Kimi-K2.6, the full NVIDIA Nemotron-3 line from Nano-30B up to Ultra-550B, GPT-OSS-120B and 20B, and a long list of Qwen models from the compact Qwen3.5-4B up to Qwen3.5-397B-A17B. Several are offered in both a standard context window and an extended 256K or 128K variant.

You can download your weights

There is an API endpoint that lets you download any checkpoint you have saved. This is the single most important line in the FAQ and it is easy to miss. A managed training service that will not give you the adapter is a service you cannot leave. Tinker will.

Cached prefill at an 80% discount

Prefill tokens that hit the cache are billed at 20% of the standard prefill rate, across every model on the list. For RL loops that repeatedly send the same long system prompt or environment description, this is the difference between a viable experiment and an unaffordable one.

A cookbook rather than a demo

Thinking Machines publishes worked recipes: prompt distillation, search tools, self-distillation fine-tuning, true-thinking scoring, plus a tutorials series. Named research groups have shipped real work on it. An MIT team used it for self-distillation fine-tuning that beat standard supervised fine-tuning on tool use, 70.6% against 63.2%. A Stanford safety group trained models whose chains of thought explain their actual behaviour, scaling up to 235B parameters. Axiom Math trains its Lean prover on it. That is a stronger evidence base than testimonials.

What does Tinker actually cost for a real run?

All prices are per million tokens and vary per model, so the honest answer is that it depends heavily on which model you pick. The spread is enormous.

At the cheap end, GPT-OSS-20B costs $0.18 for prefill, $0.45 to sample and $0.396 to train, per million tokens. Qwen3-8B is $0.195, $0.60 and $0.44. Qwen3.5-4B is $0.33, $1.005 and $0.737.

At the top, Inkling under its current limited-time 50% discount is $1.87 prefill, $4.68 sample, $5.61 train, against list prices of $3.74, $9.36 and $11.22. Kimi-K2.6 is $2.205, $5.49 and $4.84 with no discount. Qwen3.5-397B-A17B is $3.00, $7.50 and $6.60.

Checkpoint storage is charged separately at $0.10 per GB per month, which is small but not zero if you save aggressively across a long sweep.

Two things to plan around. The 50% discount is explicitly limited-time and applies only to Inkling, Inkling-Small and the Nemotron-3 family. Every third-party model on the list, Qwen, Kimi, DeepSeek and GPT-OSS, is at full price already. Budget on the list price for the first-party models and treat the current rate as a promotion, because that is what the page calls it.

Second, extended context costs roughly double. Inkling at 64K is $1.87 prefill and the 256K variant is $3.74, exactly 2x. Kimi-K2.6 goes from $2.205 at 32K to $5.15 at 128K. If your task fits in the smaller window, the smaller window is a straight 50% saving.

Why does a 27B model cost more than a trillion-parameter one?

Because the price tracks active parameters and architecture, not headline size, and the price list makes that unusually visible.

Qwen3.6-27B is a dense model. It costs $1.86 to prefill, $5.595 to sample and $4.103 to train per million tokens. Kimi-K2.6, a mixture-of-experts model many times larger by total parameter count, costs $2.205, $5.49 and $4.84. Sampling from the 27B dense model is more expensive than sampling from Kimi-K2.6.

The same generation makes the point again. Qwen3.6-35B-A3B, an MoE with 3B active parameters, costs $0.54 prefill against Qwen3.6-27B’s $1.86. Larger model on paper, 3.4 times cheaper per token, because only a fraction of it runs on any given token.

The practical lesson is to stop reading the parameter count as a price signal and read the Arch column instead. If you are picking a base model by intuition, you will systematically overpay for dense models in the 20B to 30B range.

One more asymmetry worth checking per model: training tokens are not uniformly more or less expensive than sampled ones. On Inkling, training costs $5.61 against $4.68 to sample. On Nemotron-3-Ultra-550B, training is $5.478 against $6.225 to sample, the other way round. There is no rule of thumb here that survives a model switch, so price your specific loop.

How Much Does Tinker Cost?

Tinker is usage-based with no published subscription tier and no published free credit. Prices are per million tokens, split into prefill, sample and train columns, and set per model. The floor is $0.18 per million prefill tokens on GPT-OSS-20B. Cached prefill is discounted 80% across the board. Checkpoint storage is $0.10 per GB per month. There is also a separate quantised sampling path for the Inkling family at lower rates, priced only for prefill and sampling rather than training.

The pricing does not live on the marketing page. It is in the docs, on the models page, which is a poor choice for a product where the cost model is the main purchasing question. Expect to hunt for it.

Against the field, this is competitive but not obviously cheapest, and comparison is hard because the products are shaped differently. Fireworks AI prices fine-tuning per token in bands by model size, with separate bands for models up to 16B, models from 16B to 80B, and large MoE models, and offers $1 in free credits to start. Together AI publishes per-token rates with cached input discounts across a broad model catalogue. Neither of them hands you gradient-level primitives, which is the thing you are actually buying here, so the per-token comparison undersells what Tinker does and oversells what its rivals do.

Who is Tinker Best For?

Use Tinker if you are doing research on training itself. Continual learning, reinforcement learning with custom rewards, distillation setups, anything where the training loop is the experiment rather than a step you want automated. The four primitives exist precisely so you can compose algorithms that a form-based fine-tuning service cannot express, and the published research from MIT, Stanford and Axiom Math is evidence that this actually works at scale.

Use it if you want to train small specialised models on your production traces to beat a large general model on a narrow task. Thinking Machines pitches this directly for search and data processing agents, and the cookbook has recipes for it.

Skip Tinker if you want managed fine-tuning with defaults. If your requirement is “make the model better at my support tickets” and you do not want to think about gradients, this is far more rope than you need and a simpler product will get you there faster.

Skip it if you need full-weight fine-tuning. Tinker is LoRA only. The lab argues LoRA matches full fine-tuning with the right setup and publishes research to that effect, but if your work requires updating base weights, that is a hard stop rather than a trade-off.

And skip it if you need to know your bill in advance. There is no cap, no plan, and no free tier published anywhere. Everything is metered.

Best Tinker Alternatives

Fireworks AI offers fine-tuning priced per token in bands by parameter size, with a separate band for large MoE models, plus $1 in free credits and postpaid billing. It is the closer commercial comparison on price structure, though it sells a managed pipeline rather than training primitives.

Together AI publishes per-token pricing with cached input discounts across a wide open-model catalogue and offers fine-tuning alongside inference. Pick it if you want one vendor for both training and serving, which Tinker does not currently try to be.

Running it yourself is the real alternative for anyone with GPU access. LoRA on your own hardware costs nothing per token and gives you total control. What you give up is exactly what Tinker sells: scheduling, distributed orchestration, reliability, and not having to become a cluster operator. If you already have the cluster and the operator, the case for Tinker weakens sharply.

Final Verdict: Is Tinker Worth It?

This is the most credible managed training API we have looked at. The API design is restrained in a way that suggests real taste: four functions that compose, rather than fifty that cover cases. The model roster is broad and current. Cached prefill at 80% off is exactly the discount an RL researcher needs. And you can download your checkpoints, which means adopting it is not a one-way door.

The reservations are commercial rather than technical. Pricing is buried in the docs when it should be a link in the nav, there is no free tier to evaluate on, the headline discount on the first-party models is explicitly temporary, and the per-model price spread is wide enough that a careless base-model choice can cost you several times more than a careful one for no quality gain. Read the models table before you write a line of code, pick on the Arch column rather than the parameter count, and this is an easy recommendation for anyone whose work is the training loop itself.